-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodel.py
232 lines (203 loc) · 7.06 KB
/
model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
#! /usr/bin/env python
# coding=utf-8
# /************************************************************************************
# ***
# *** File Author: Dell, Thu Sep 20 21:42:14 CST 2018
# ***
# ************************************************************************************/
import os
import torch
import torch.nn as nn
from torchvision import transforms
def image_to_tensor(image):
"""
return 1xCxHxW tensor
"""
transform = transforms.Compose([
transforms.Resize(size=(512)),
transforms.ToTensor(),
])
t = transform(image)
t.unsqueeze_(0)
return t
def image_from_tensor(tensor):
"""
tensor format: 1xCxHxW
"""
transform = transforms.Compose([
transforms.ToPILImage(),
])
tensor.squeeze_(0)
return transform(tensor)
def mean_std(feat, eps=1e-5):
# set eps to avoid divide-by-zero.
size = feat.size()
assert (len(size) == 4)
N, C = size[:2]
feat_var = feat.view(N, C, -1).var(dim=2) + eps
feat_std = feat_var.sqrt().view(N, C, 1, 1)
feat_mean = feat.view(N, C, -1).mean(dim=2).view(N, C, 1, 1)
return feat_mean, feat_std
def feat_normal(c_feat, s_feat, eps=1e-5):
assert (c_feat.size()[:2] == s_feat.size()[:2])
size = c_feat.size()
s_mean, s_std = mean_std(s_feat)
c_mean, c_std = mean_std(c_feat)
c_std.add_(eps)
nc = (c_feat - c_mean.expand(size)) / c_std.expand(size)
return nc * s_std.expand(size) + s_mean.expand(size)
vgg_decoder = nn.Sequential(
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 256, (3, 3)),
nn.ReLU(),
nn.Upsample(scale_factor=2, mode='nearest'),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 256, (3, 3)),
nn.ReLU(),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 256, (3, 3)),
nn.ReLU(),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 256, (3, 3)),
nn.ReLU(),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 128, (3, 3)),
nn.ReLU(),
nn.Upsample(scale_factor=2, mode='nearest'),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(128, 128, (3, 3)),
nn.ReLU(),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(128, 64, (3, 3)),
nn.ReLU(),
nn.Upsample(scale_factor=2, mode='nearest'),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(64, 64, (3, 3)),
nn.ReLU(),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(64, 3, (3, 3)),
)
vgg_encoder = nn.Sequential(
nn.Conv2d(3, 3, (1, 1)),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(3, 64, (3, 3)),
nn.ReLU(), # relu1-1
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(64, 64, (3, 3)),
nn.ReLU(), # relu1-2
nn.MaxPool2d((2, 2), (2, 2), (0, 0), ceil_mode=True),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(64, 128, (3, 3)),
nn.ReLU(), # relu2-1
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(128, 128, (3, 3)),
nn.ReLU(), # relu2-2
nn.MaxPool2d((2, 2), (2, 2), (0, 0), ceil_mode=True),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(128, 256, (3, 3)),
nn.ReLU(), # relu3-1
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 256, (3, 3)),
nn.ReLU(), # relu3-2
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 256, (3, 3)),
nn.ReLU(), # relu3-3
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 256, (3, 3)),
nn.ReLU(), # relu3-4
nn.MaxPool2d((2, 2), (2, 2), (0, 0), ceil_mode=True),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(256, 512, (3, 3)),
nn.ReLU(), # relu4-1, this is the last layer used
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU(), # relu4-2
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU(), # relu4-3
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU(), # relu4-4
nn.MaxPool2d((2, 2), (2, 2), (0, 0), ceil_mode=True),
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU(), # relu5-1
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU(), # relu5-2
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU(), # relu5-3
nn.ReflectionPad2d((1, 1, 1, 1)),
nn.Conv2d(512, 512, (3, 3)),
nn.ReLU() # relu5-4
)
def encoder_load(model):
encoder = vgg_encoder
if os.path.exists(model):
encoder.load_state_dict(torch.load(model))
layers = list(encoder.children())
encoder = nn.Sequential(*layers[:31])
return encoder
def decoder_load(model):
decoder = vgg_decoder
if os.path.exists(model):
decoder.load_state_dict(torch.load(model))
return decoder
class StyleNet(nn.Module):
def __init__(self, encoder, decoder):
super(StyleNet, self).__init__()
enc_layers = list(encoder.children())
self.enc_1 = nn.Sequential(*enc_layers[:4]) # input -> relu1_1
self.enc_2 = nn.Sequential(*enc_layers[4:11]) # relu1_1 -> relu2_1
self.enc_3 = nn.Sequential(*enc_layers[11:18]) # relu2_1 -> relu3_1
self.enc_4 = nn.Sequential(*enc_layers[18:31]) # relu3_1 -> relu4_1
self.decoder = decoder
self.mse_loss = nn.MSELoss()
# fix the encoder
for name in ['enc_1', 'enc_2', 'enc_3', 'enc_4']:
for param in getattr(self, name).parameters():
param.requires_grad = False
# extract relu1_1, relu2_1, relu3_1, relu4_1 from input image
def encode_layers(self, input):
results = [input]
for i in range(4):
func = getattr(self, 'enc_{:d}'.format(i + 1))
results.append(func(results[-1]))
return results[1:]
# extract relu4_1 from input image
def encode(self, input):
for i in range(4):
input = getattr(self, 'enc_{:d}'.format(i + 1))(input)
return input
def content_loss(self, input, target):
assert (input.size() == target.size())
assert (target.requires_grad is False)
return self.mse_loss(input, target)
def style_loss(self, input, target):
assert (input.size() == target.size())
assert (target.requires_grad is False)
input_mean, input_std = mean_std(input)
target_mean, target_std = mean_std(target)
return self.mse_loss(input_mean, target_mean) + self.mse_loss(input_std, target_std)
def forward(self, content, style, alpha=1.0):
assert 0 <= alpha <= 1
s_feats = self.encode_layers(style)
c_feat = self.encode(content)
t = feat_normal(c_feat, s_feats[-1])
t = alpha * t + (1 - alpha) * c_feat
g_t = self.decoder(t)
g_t_feats = self.encode_layers(g_t)
# g_t_feats[-1] ==> last row ...
loss_c = self.content_loss(g_t_feats[-1], t)
loss_s = self.style_loss(g_t_feats[0], s_feats[0])
for i in range(1, 4):
loss_s += self.style_loss(g_t_feats[i], s_feats[i])
return loss_c, loss_s
def style_transfer(encoder, decoder, content, style, alpha=1.0):
assert (0.0 <= alpha <= 1.0)
c_f = encoder(content)
s_f = encoder(style)
feat = feat_normal(c_f, s_f)
feat = feat * alpha + c_f * (1 - alpha)
return decoder(feat)